“Our company only has about ten people — do we even need an enterprise knowledge base?”
I get asked this a lot. The people asking are usually already using AI customer service, or about to try it.
Conclusion first: yes, you do. But not in the way you are probably imagining.
The moment most bosses hear “knowledge base”, the picture that comes to mind is buying a system and uploading every file the company has. Do that, and it basically dies in the drawer. Dump files in for three months and nobody ever opens them.
The reason is simple. A knowledge base is not a file cabinet — it is a set of answers.
What staff ask is “can this be returned”, not “find the after-sales policy document”. Those two sentences are very different: one is getting work done, the other is looking up reference material.
The 5 steps below are the fixed order we use when we deploy for small businesses. It is not big, it is not expensive, and one person can produce a first version in three to five days.
Step 1: First circle out the 50 questions that keep getting asked
Do not buy software first. Flip through the chat records first.
Customer service backend, after-sales groups, the boss’s own WeChat — copy down the questions that have been asked in the last three months. Get to 50, then sort them by how often each appears.
The top 20 questions usually carry eight tenths of the inquiry volume. How much is shipping, how many days to arrive, can it be returned, how do I fix it when it breaks, do you issue an invoice.
Once this step is done, you hold a real question list in your hand.
Do not underestimate it. It is worth more than any industry template.
Copying out the list looks dumb. Half a day gets it done.
Step 2: Write the answers in plain language
One question, one answer. Write short. Write specific.
Do not hoard them and write in bulk later.
“Shipping is 8 yuan, free at 99 and above, Xinjiang and Tibet add 15”
“Seven-day no-reason return, returns accepted even after opening the package, but not for customized items”
Do not write “refer to Chapter 3 of the company after-sales management policy”. Nobody works by chapter, and the AI does not recognize chapters either.
The test is simple: can a new hire, holding this answer, reply to the customer directly? Yes, then it is a passing answer.
Step 3: Split the content into three layers
This is the step I see missed most often. If all content sits mixed together, the AI cannot tell which lines may be said and which may not.
- The wording layer: unified external scripts. Prices, delivery times, return and exchange standards.
- The rules layer: red lines. Things that must not be promised, topics that must not be answered.
- The evidence layer: sources. Original platform rules, national standard clauses, contract terms.
Many people skip the third layer. They find it a pain.
But the moment a customer pushes hard, or the platform comes to inspect, the one who can produce the source does not lose.
Example: the standard basis for AI customer service is the national standard GB/T 47746—2026, which took effect on 1 September 2026. It sets out 61 requirements (48 “shall” items, 4 “should” items, 9 “may” items), of which 5 are one-vote-veto items, covering scenarios such as automatic handover to a human.
Put clauses like these into the evidence layer, and the AI has backing when it answers.
Step 4: When it cannot answer, hand over to a human
This step decides whether going live saves trouble or creates it.
Draw a line for the AI: answers that are in the base, it gives; answers that are not in the base, hand straight to a human.
Do not count on it to “guessed cleverly”. The moment the AI guesses, it is making things up. A made-up shipping rate, a made-up warranty period — the customer screenshots it and sends it out, and what you lose is more than what you saved.
The handover-to-human judgment goes into the system, not into the employee handbook. What is written in a handbook is ignored by everyone within three days.
The system intercepts; people do not watch every day.
Step 5: Iterate weekly against the “wrong answers list”
Going live is not the finish line.
Keep a notebook, or a sheet, dedicated to the questions the AI answered wrong. Each week, pick a few and add them to the base.
Hold this for two months and you will see the inquiry distribution change — the same questions keep being asked, but nobody comes to you anymore; the AI has answered them all.
Two pitfalls, up front
Pitfall 1: buy the tool first, organize the content later.
Wrong order. If the content is not sorted, the most expensive system on the market just idles. The other way around — get the answers to 50 questions organized, and it runs fine on the simplest tool.
Pitfall 2: organized once, nobody maintains it.
A knowledge base goes stale. Shipping prices go up, promotion rules change, a new product launches. Assign one person to spend thirty minutes a week updating it. Thirty minutes is far cheaper than mopping up after something goes wrong.
One last line
With an enterprise knowledge base, the hard part is never the technology.
The hard part is whether you are willing to sit down and write the answers to those 50 questions out, one by one, instead of leaving them in your employees’ heads.
The day that is written clearly is the day your AI customer service can truly be called live. That one step has no shortcut.
#enterprise-knowledge-base #AI-customer-service #SME-digitalization
If you are organizing your own company’s knowledge base and are stuck at some step, tell me in the comments which step — I will pick a few and write them up in the next article.
Who we are: we help small and micro businesses build enterprise knowledge bases and land AI customer service. We turn the wording, rules and scripts scattered in employees’ heads into a base that the AI can look up and that is traceable; then we add red-line interception and handover-to-human judgment to the AI customer service.